NVIDIA H200 NVL compatible servers: card counts, NVLink bridges, PSUs and inlet limits

Which servers list the H200 NVL

The NVIDIA H200 NVL runs in servers whose maker lists it for that model, with a specified number of cards, NVLink bridges, power supplies and cooling. In the makers' documents we read on 24 October 2026, seven servers list it: Dell's PowerEdge XE7740, Cisco's UCS C845A M8, HPE's ProLiant Compute DL380a Gen12, Lenovo's ThinkSystem SR650a V4 and SR675 V3, and Supermicro's SYS-212GB-NR and SYS-521GE-TNRT. Two cards fit 2U servers, four fit some 2U and every 4U model here, and eight need 3U to 5U with two NVLink domains of four.

Dell PowerEdge XE7740 technical guide, Rev. A04, April 2026; Cisco UCS C845A M8 spec sheet, Rev. A.11, 11 May 2026, and data sheet of 24 August 2026; HPE ProLiant Compute DL380a Gen12 QuickSpecs (a00047453enw); Lenovo Press LP2128 (20 October 2026), LP1611 (8 October 2026) and LP1944 (7 November 2025) with Lenovo ServerProven; Supermicro product pages. All read on 24 October 2026. "Not stated" means the document we read gives no H200-specific value.

The table lists servers, not our products: we supply the H200 NVL and its bridges for servers like these, and build our own AI servers to order around the card. Lenovo's ServerProven entry for the card names only the SR650a V4 and the SR675 V3, and states that the H200 is "not supported on SR650 V4". Supermicro's pages for the AS-4125GS-TNRT and the SYS-522GA-NRT, both 8-GPU systems, list RTX PRO cards but no H200 NVL. Dell's technical guide for the XE7745 (Rev. A09, September 2026) lists up to eight H200 NVL, double-width at PCIe Gen5 x16, with a maximum ambient of 30 °C, and MLPerf® Inference: Datacenter v5.0, closed division, Llama 2 70B, offline scenario, entry 5.0-0017 ran eight H200 NVL in an XE7745 (retrieved from mlcommons.org on 24 September 2026, result verified by MLCommons Association). The MLPerf name and logo are registered and unregistered trademarks of MLCommons Association in the United States and other countries. All rights reserved. Unauthorized use strictly prohibited. See www.mlcommons.org for more information. For any model, ask the maker for the current list before you order.

Two cards: 2U servers and their limits

Lenovo's SR650a V4, a 2U server with one or two Xeon 6 processors, lists "2x NVIDIA H200 NVL 600W GPUs" in its product guide of 20 October 2026, in the front GPU bay, which holds up to four double-wide GPUs in total. Lenovo's guide gives no bridge layout for the H200 NVL in this server. Cisco's C845A M8 takes two cards in slots S8 and S7 with one 2-way bridge, and HPE's DL380a Gen12 takes two double-wide cards with five power supplies, two for the system board and three for the GPUs. Two cards with a 2-way bridge give 282 GB in one NVLink domain, which our guide to the 2× H200 NVL server sizes.

Four cards: one 4-way bridge or two 2-way

Cisco documents three ways to fit four cards: slots S8 to S5 with one 4-way bridge, the same slots with two 2-way bridges, or two pairs in slots S8, S7, S4 and S3, one 2-way bridge in each of the two GPU blocks. Dell uses one 4-way bridge for four cards in the XE7740. Supermicro's SYS-212GB-NR takes four double-wide cards in 2U, with one Xeon 6 processor up to 350 W, 16 memory slots and an optional NVLink bridge.

Where the cards sit matters as much as the bridge. NVIDIA's reference architecture for the H200 NVL states that "pairing of GPU cards under the same CPU socket is best", and Dell writes: "Although it is possible to bridge H200 GPUs with 4-way NVLink under different CPUs, it is not recommended." In the XE7740, slots 21 to 28 connect to the first processor and 29 to 36 to the second, and Dell's slot priority fills 25, 27, 29 and 31 first, two under each processor, so a four-card bridged layout is worth confirming with Dell. Our guide to the 4× H200 NVL server covers what one domain of 564 GB runs. We supply the cards and bridges for each layout; send us the server model and its free slots, and we check bridge, cable, power supply and inlet limit before you order.

Eight cards: two domains of four

Every eight-card server here forms two NVLink domains of four, joined over PCIe. Dell's guide states: "For four GPUs, one 4-way bridge is needed, while for eight DW H200 GPU cards, two bridges are required." Cisco's Table 17 allows eight cards in slots S8 to S1 as two 4-way bridges, one per GPU block (S5 to S8 and S1 to S4), or four 2-way bridges, and six cards as three 2-way bridges or one 4-way plus one 2-way. Cisco adds that cards start in slot 8 and are added two adjacent cards at a time.

HPE's DL380a Gen12 allows 2, 4, 8 or 10 double-wide GPUs, but its QuickSpecs state that the "10DW configuration does not support GPU NVL bridges", so eight is the bridged maximum. Lenovo's SR675 V3 takes eight double-wide cards at 600 W each in its 8-DW base, and Lenovo's guide for the card lists the server with 5th Gen AMD EPYC processors only. Supermicro's SYS-521GE-TNRT takes eight double-wide cards behind dual-root PCIe 5.0 switches. How a model is split over two domains is in our guide to what an 8× H200 NVL server runs.

Bridges are separate parts

NVIDIA's product brief for the card (PB-12128-001_v01, April 2025) states that "the NVIDIA H200 NVL uses a single wide NVLink bridge per card". Every maker sells the bridges under its own part number, and HPE states that its "4-way and 2-way Bridge for H200 NVL can only be supported with H200 NVL GPUs", so H100 NVL bridges do not carry over.

Cisco C845A M8 spec sheet Rev. A.11; HPE DL380a Gen12 QuickSpecs; Lenovo Press LP1944; Dell XE7740 technical guide Rev. A04; read on 24 October 2026. Dell's guide names the 4-way bridge for the H200 NVL and the 2-way bridge for the H100 NVL; it gives no bridge part numbers in the sections we read.

Order the bridges with the cards. Dell warns that removing a bridge without its two tools, used on both sides at once, can damage the NVLink interface, and Cisco also lists a separate removal kit. The bridge mechanics and bandwidth per domain are in our article on the H200 NVL NVLink bridge, 2-way or 4-way.

Power supplies and 16-pin cables

Each card takes up to 600 W through one 16-pin connector, and NVIDIA's product brief sets 600 W as the default cap with 200 W as the minimum; a card whose cable signals less than its default cap does not boot. The makers specify the power side in detail.

• Dell XE7740. Up to eight supplies in two zones, 1+1 for the processors and 3+3 for the GPUs. The 3,200 W Titanium supply delivers 3,200 W at 220.1 to 240 V and 2,900 W at 200 to 220 V, and Dell's power matrix marks 2,400 W supplies with eight H200 NVL as not supported in the 4+0, 3+1 and 4+4 configurations, while 8+0 and 7+1 are listed as supported.
• HPE DL380a Gen12. "2400W or 3200W PSUs must be selected" with the H200 NVL; five supplies for two or four double-wide GPUs, eight for eight or ten; C19 or C19-C20 cords only; one 16-pin cable kit for every two double-wide GPUs.
• Cisco C845A M8. Four 3.2 kW Titanium supplies with N+1 redundancy, all four installed, C20 inlets, 100 to 240 V input.
• Lenovo. Cables are server-specific: a 320 mm 16-pin cable for the SR650a V4 and a 235 mm cable for the SR675 V3, not included with the card's option part. The SR675 V3 takes up to four 1,800, 2,400 or 2,600 W supplies at 220 V.
• Supermicro. Three 2,000 W supplies in 2+1 for the 2U SYS-212GB-NR, four 2,700 W supplies at 200 to 240 V for the 5U SYS-521GE-TNRT.

Eight cards at 600 W draw 4.8 kW before processors, memory and fans, which puts an eight-card server on a three-phase feed position, as our guide to GPU rack power and cooling works out. For a card going into a server you already run, our H200 NVL in an existing server checklist covers cable, sense pins and slot power.

Inlet temperature: 30 °C in some chassis

The H200 NVL is passive and depends on the server's fans. Dell's XE7740 guide lists the ambient limit for configurations with eight GPUs per card type: 30 °C for the H200 NVL, against 35 °C for the H100 NVL, L40S and L4. The general range of the server is 10 to 35 °C up to 900 m, falling by 1 °C per 300 m above that. Supermicro gives 10 to 35 °C for both of its systems without a card-specific value, HPE's QuickSpecs refer to a separate ASHRAE page, and Cisco's spec sheet gives no inlet limit for the card. A data centre that runs cold aisles at 27 °C has margin for the Dell limit; a server room at 32 °C does not.

Host processors and memory by maker

Memory rules differ more than any other line in these documents.

Makers' documents as in the first table, read on 24 October 2026; NVIDIA AI Enterprise reference architecture (compute node) for the 128 GB per GPU minimum.

NVIDIA's reference architecture asks for at least 128 GB per GPU and seven cores per GPU plus two per GPU for the operating system, and NVIDIA's configuration guide for NVIDIA-Certified Systems asks for at least twice the GPU memory. For eight cards that ranges from 1 TB to Cisco's 4 TB. Dell also warns that the configurations it has not validated "may result in GPU performance degradation, PCIe fatal errors, or system boot issues".

NVIDIA-Certified systems and where the cards come from

NVIDIA's H200 page names "MGX H200 NVL partner and NVIDIA-Certified Systems with up to 8 GPUs". Certification matters for the five-year NVIDIA AI Enterprise subscription included with every H200 NVL: NVIDIA's licensing guide states that the software "can be installed on the NVIDIA Certified Systems in which the GPUs are installed, but not on a different server or instance". Some makers also tie support to their own card part. Cisco asks for GPUs bought from Cisco, because its management firmware needs a Cisco-specific SBIOS ID, and HPE and Lenovo sell the card under their own part numbers. For a server under the maker's support contract, check which card part it accepts before you order.

What we supply

We supply the NVIDIA H200 NVL with its 2-way and 4-way NVLink bridges, as cards for a server you already run or in AI servers built to order, including our training node with eight cards on two 4-way bridges, on one EU contract and invoice with manufacturer warranty handled through us. Each card includes its five-year NVIDIA AI Enterprise subscription. Name your server model, its free slots and the rack's inlet temperature, and our team replies within one business day with the cards, bridges and cables it needs, or with a server configuration if your chassis is not on a list.

FAQ